Lizandro D. Solano-Quinde

Unstructured grid applications on GPU: performance analysis and improvement

Performance of applications running on GPUs is mainly affected by hardware occupancy and global memory latency. Scientific applications that rely on analysis using unstructured grids could benefit from the high performance capabilities provided by GPUs, however, its memory access pattern and algorithm limit the potential benefits. In this paper we analyze the algorithm for unstructured grid analysis on the basis of hardware occupancy and memory access efficiency. In general, the algorithm can be divided into three stages: cell-oriented analysis, edge-oriented analysis and information update, which present different memory access patterns. Based on the analysis we modify the algorithm to make it suitable for GPUs. The proposed algorithm aims for high hardware occupancy and efficient global memory access. Finally, through implementation we show that our design achieves up to 88 times speedup compared to the sequential CPU version.

Module Prototype for Online Failure Prediction for the IBM Blue Gene/L

The growing complexity of scientific applications has led to the design and deployment of large-scale parallel systems. The IBM Blue Gene/L can hold in excess of 200 K processors and it has been designed for high performance and reliability. However, failures in this large-scale parallel system are a major concern, since it has been demonstrated that a failure will significantly reduce the performance of the system.

Coarse grain computation-communication overlap for efficient application-level checkpointing for GPUs

Graphics Processing Units (GPUs) are increasingly used to solve non-graphical scientific problems. However, it has been shown that the reliability of the GPUs is a concern because of the occurrence of the soft and hard errors. The checkpoint/restart is the most commonly used technique to achieve fault tolerance in the presence of failures. This work present an application-level checkpoint scheme for systems composed of GPUs. Our scheme exploits the benefits of the divide-and-conquer technique and of the communication-computation overlapping to improve the execution time and checkpoint overhead. By dividing the problem and checkpointing in n subprocesses, we show that our scheme improves the checkpoint overhead by a factor of n. We also show that dividing the problem with finer granularity is not beneficial.

Techniques for the parallelization of unstructured grid applications on multi-GPU systems

Currently the set of scientific applications suitable for running on GPUs has increased due to the computational power of GPUs and the availability of programming languages that make more approachable writing scientific applications for GPUs. However, as the size of the problems increases, the global memory of GPUs becomes a limitation for running applications. Multi-GPU systems can potentially make memory limited problems tractable by dividing the data and computation among several GPUs. Parallel execution is seriously limited by the (i) application data dependencies, and (ii) data transfers among GPUs. In this paper we analyze the potential for parallelization of unstructured grid applications based on the data dependencies of the algorithm and the amount of data communication required. Due to data dependencies and the required communication, data and task parallelization techniques present different communication overheads and computing devices utilization. Based on this analysis we propose a scheme that takes advantage of data and task parallelism and reduces the communication overhead through computation-communication overlap. Our OpenCL implementation reduces the communication overhead by 38%, and, for comparison purposes, a two GPU implementation provides almost a five-fold increase in performance as compared to a CPU implementation.

Characterizing artifacts in RR stress test time series

Electrocardiographic stress test records have a lot of artifacts. In this paper we explore a simple method to characterize the amount of artifacts present in unprocessed RR stress test time series. Four time series classes were defined: Very good lead, Good lead, Low quality lead and Useless lead. 65 ECG, 8 lead, records of stress test series were analyzed. Firstly, RR-time series were annotated by two experts. The automatic methodology is based on dividing the RR-time series in non-overlapping windows. Each window is marked as noisy whenever it exceeds an established standard deviation threshold (SDT). Series are classified according to the percentage of windows that exceeds a given value, based upon the first manual annotation. Different SDT were explored. Results show that SDT close to 20% (as a percentage of the mean) provides the best results. The coincidence between annotators classification is 70.77% whereas, the coincidence between the second annotator and the automatic method providing the best matches is larger than 63%. Leads classified as Very good leads and Good leads could be combined to improve automatic heartbeat labeling.

Minimizing the power consumption in Raspberry Pi to use as a remote WSN gateway

The reliability of Raspberry Pi minicomputer as a remote gateway in a wireless sensor network requires an analysis of power consumption. Due to the limited energy conditions, when a station is powered by a photovoltaic system, it is essential to reduce its energy consumption to maximize the lifetime of the network. This paper presents a comparative analysis of Raspberry Pi power consumption profile under multiple settings: reduced hardware elements, installation of different operating systems and, downclock CPU (700 MHz to 100 MHz), SDRAM (400 MHz to 100 MHz) and GPU (250 MHz to 50 MHz) units. Based on the results, the energy consumption is reduced up to 20% which is equivalent to a saving of 0.25 W. The minimum Raspberry Pi power consumption reached is 0.97 W.

Automatic Parallelization of GPU Applications Using OpenCL

Graphics Processing Units (GPUs) have been successfully used to accelerate scientific applications due to their computation power and the availability of programming languages that make more approachable writing scientific applications for GPUs. However, since the programming model of GPUs requires offloading all the data to the GPU memory, the memory footprint of the application is limited to the size of the GPU memory. Multi-GPU systems can make memory limited problems tractable by parallelizing the computation and data among the available GPUs. Parallelizing applications written for running on single-GPU systems can be done (i) at runtime through an environment that captures the memory operations and kernel calls and distributes among the available GPUs, and (ii) at compile time through a pre-compiler that transforms the application for decomposing the data and computation among the available GPUs. In this paper we propose a framework and implement a tool that transforms an OpenCL application written to run on single-GPU systems into one that runs on multi-GPU systems. Based on data dependencies and data usage analysis, the application is transformed to decompose data and computation among the available GPUs. To reduce the data transfer overhead, computation-communication overlapping techniques are utilized. We tested our tool using two applications with different data transfer requirements, for the application with no data transfer requirements, a linear speedup is achieved, while for the application with data transfers, the computation-communication overlapping reduces the communication overhead by 40%.

Multi-GPU implementation of the Horizontal Diffusion method of the Weather Research and Forecast Model

The Weather Research and Forecasting (WRF), a next generation mesoscale numerical weather prediction system, has a considerable amount of work regarding GPU acceleration. However, the amount of works exploiting multi-GPU systems is limited. This work constitutes an effort on using GPU computing over the WRF model and is focused on a computationally intensive portion of the WRF: the Horizontal Diffusion method. Particularly, this work presents the enhancements that enable a single-GPU based implementation to exploit the parallelism of multi-GPU systems. The performance of the multi-GPU and single-GPU based implementations are compared on a computational domain of 433x308 horizontal grid points with 35 vertical levels, and the resulting speedup of the kernel is 3.5x relative to one GPU. The experiments were carried out on a multi-core computer with two NVIDIA Tesla K40m GPUs.

Evaluating reliability of ultrashort heart rate variability parameters in metabolic syndrome subjects

Heart rate variability (HRV) analysis is barely employed in healthcare environments mainly because of the lack of standard values determining the sympathovagal balance and the difficulty to register RR stationary series. Recent studies have proposed the use of shorter HRV series. For this work, we use a public metabolic syndrome subjects database retrieved during oral glucose tolerance test. In order to explore ultra-short HRV measures reliability we employ an autoregressive model using Burg method, such that short RR sequences can be evaluated while maintaining a good frequency resolution. RR, SD, rMSSD, LF, HF, LFn and LF/HF were computed for different RR sequences (10 min, 5 min, 1 min, 30 s, 10 s). To evaluate the reliability we used the intraclass correlation coefficient (ICC). Additionally, we compared the sympathovagal balance parameters (LFn, LF/HF) among the stages (basal and 30 min). Considering 10 min long registers as references, parameters obtained from 5 min long series present ICC values above 0.78 for all cases. One min long registers present ICC values above 0.70 only for temporal parameters in both RR series and rMSSD. By comparing LFn and LF/HF parameters among the basal state and 30 min, we observed a significant increase of the sympathetic tone (p <; 0.05). However, these differences are important only for 10 and 5 min series. In general, we observe that temporal parameters exhibit higher reliability than those the spectral ones. Nonetheless, registers duration below one min do not present adequate results for the spectral parameters in this work.